Synopsis

This research introduces the Physics-Informed Variational Autoencoder (PI-VAE), a novel framework designed to reconstruct cosmological fields and infer fundamental parameters directly from high-dimensional simulation data. By embedding physical constraints—such as the matter power spectrum and mass conservation—directly into the training objective, the model overcomes the "shortcut learning" limitations of traditional deep learning.

Using multi-channel data (Gas Mass, Velocity, and Magnetic Fields) from the CAMELS dataset, PI-VAE achieves near-perfect inference for matter density (Ωm​) with an R2 score of 0.957 and a robust R2 of 0.732 for fluctuation amplitude (σ8​). Notably, the model improves physics fidelity by 98.6% over vanilla VAEs, ensuring that reconstructed maps of the Cosmic Web are both visually sharp and scientifically valid. This work provides a foundation for scaling AI-driven cosmological analysis to 3D data and real-world telescope surveys.

Relevance of the Topic

(write your relevancy here)

Future Research/Scope

(write your future scope here)

Skills Learned

(write your Skills acquired here)

Relevant courses to the topic

• (Course list here)

Reading List

• (reading list here)